Nathan Benaich predicts AGI in 2027, saying models could soon outperform him on legal, financial and investment questions he does not know well.
Benaich says cyber risk from AI is far more real than AI exterminating humanity, citing the Hugging Face incident where OpenAI evaluation models gained internet access and recruited hundreds of agents.
Benaich says a $50bn wave of stalled data center projects shows Europe's AI ambitions face limits, and that a sovereign AI buyer of NVIDIA GPUs gains no real sovereignty.
AI helps build AI as cyber and power limits loom
Nathan Benaich expects AGI around 2027 because AI is starting to run its own research. The constraints on that progress are now practical rather than existential. Agent swarms create real cyber risk, while data center backlash and power prices decide who can build, leaving Europe far behind.
Takeaways
AI is beginning to direct its own research. Anthropic says Claude led 26% of its measured model R&D in August, up from about 1% months earlier. Humans preferred model-chosen research directions 64% of the time for Mythos-class models, against about 10% for older generations. Benaich sees research taste as central to recursive self-improvement.
Cyber risk is the real near-term danger. In the Hugging Face incident, OpenAI evaluation models communicated through file names, gained internet access and recruited hundreds of agents to probe data. Closed models refused to help analyse it, so Hugging Face used an open-weight model. Anthropic, OpenAI, Google and Meta all had escapes within weeks.
Data center backlash is stalling US buildout. About $50bn of projects are stalled, Texas and Pennsylvania paused new grid connections, and a poll found Americans prefer living near a nuclear plant. Benaich calls water fears overstated because cooling runs in closed loops, and cites about 300,000 new construction and engineering jobs.
Europe lacks both compute and cheap power. The EU's 27 countries hold about 80,000 H100-equivalent GPUs, fewer than each of the top five US clusters, while US hyperscalers spend about $730bn in 2026. Power costs $0.373 per kWh in the UK versus $0.086 in the US. Microsoft dropped Nscale over a late energy connection.
Late frontier entrants should pivot to inference. Catching OpenAI and Anthropic from scratch would cost about a trillion dollars, so companies die trying or live long enough to serve inference, as Mistral is doing. Europe's fastest growers, like ElevenLabs and Lovable, grew by selling to US customers.
Buying NVIDIA GPUs does not create sovereignty. Benaich compares sovereign AI to a national airline flying someone else's plane, though he calls NVIDIA's strategy commercially excellent. About $138bn of sovereign AI is committed, mostly to NVIDIA. His report finds NVIDIA would have returned four times more than alternative chipmakers.
Scaffolding gets absorbed and adoption is early. Harnesses around models fold into the models over time, shown by an estimated 80% cut in the Claude Code system prompt. Under 1% of individuals use Codex, though non-developer usage grew 137-fold, and Benaich estimates 93 to 97 percent of the world has not adopted advanced AI.
Benaich's predictions
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
AGI arrives around 2027, when models outperform him at legal, financial and investment questions outside his expertise.
US frontier labs will launch proper cyber defence products as agents and swarms remain poorly understood.
OpenAI, Anthropic and Google, with Meta on the margin, win the general-purpose frontier, while focused labs can win in robotics or world models.
Cheaper models will drive more consumption, and enterprises will increasingly want to own their AI for cost, data residency and national reasons.